GSPlane: Concise and Accurate Planar Reconstruction via Structured Representation

TL;DR

GSPlane addresses the challenge of reconstructing accurate planar geometry within Gaussian Splatting representations by injecting 2D planar priors from segmentation and normals, re-parameterizing planar Gaussians on planes with a weighted basis formulation, and employing a Dynamic Gaussian Re-classifier to curb misclassifications. A mesh layout refinement step leverages these priors to produce topology-consistent planar regions with fewer vertices, while Supportive Plane Correction maintains plane integrity for object decoupling on planes. Evaluations on ScanNetV2 and Tanks & Temples show improved planar geometry and reduced mesh complexity without sacrificing rendering quality, across multiple baselines. The work enables more editable, plane-aware 3D reconstructions suitable for digital twins and physical simulations.

Abstract

Planes are fundamental primitives of 3D sences, especially in man-made environments such as indoor spaces and urban streets. Representing these planes in a structured and parameterized format facilitates scene editing and physical simulations in downstream applications. Recently, Gaussian Splatting (GS) has demonstrated remarkable effectiveness in the Novel View Synthesis task, with extensions showing great potential in accurate surface reconstruction. However, even state-of-the-art GS representations often struggle to reconstruct planar regions with sufficient smoothness and precision. To address this issue, we propose GSPlane, which recovers accurate geometry and produces clean and well-structured mesh connectivity for plane regions in the reconstructed scene. By leveraging off-the-shelf segmentation and normal prediction models, GSPlane extracts robust planar priors to establish structured representations for planar Gaussian coordinates, which help guide the training process by enforcing geometric consistency. To further enhance training robustness, a Dynamic Gaussian Re-classifier is introduced to adaptively reclassify planar Gaussians with persistently high gradients as non-planar, ensuring more reliable optimization. Furthermore, we utilize the optimized planar priors to refine the mesh layouts, significantly improving topological structure while reducing the number of vertices and faces. We also explore applications of the structured planar representation, which enable decoupling and flexible manipulation of objects on supportive planes. Extensive experiments demonstrate that, with no sacrifice in rendering quality, the introduction of planar priors significantly improves the geometric accuracy of the extracted meshes across various baselines.

Figures (9)

• Figure 1: We introduce GSPlane, which adopts 2D planar priors to constrain planar Gaussian distributions on corresponding planes. The structured representation for planes not only empowers the layout refinement of the mesh, resulting in topological correctness with notable less vertices, but also demonstrates the potential in decoupling objects and sealing contact regions on supportive planes.
• Figure 2: Pipeline of GSPlane. Given a set of posed images as input, our method first extracts 2D planar priors from each view, align them with point cloud to obtain plane distributions in 3D space, and re-parameterize the related coordinates of Gaussians. During training, our Dynamic Gaussian Re-classifier continues to correct false-positive planar gaussian by reverting their representation back to $xyz$. The layout of the mesh extracted from training will also be refined with the optimized planar information from the structured representation.
• Figure 3: Visualizations of the mesh performance on both indoor and outdoor scenes. We provide comparisons on four baseline methods. It can be seen from the refined normal map and wireframes that our method can reduce the number of vertices by large margin, while maintaining consistent normal and topology across different planes. More examples can be found in Appendix.
• Figure 4: Ablation study results on GSPlane. The left table shows the effectiveness of different modules in GSPlane, and the right table compares our structured representation with off-the-shelf normal map supervision for mesh geometry reconstruction.
• Figure 5: Visualizations of Supportive Plane Correction. When running SPC, the object boundaries are excluded from plane reconstruction, leading to an intact plane with complete shape like in reality. The objects are decoupled from the supportive plane surface, and can be further moved or manipulated freely.